In this paper, I review approaches for acquiring hierarchical knowledge to improve the effectiveness of planning systems. First I note some benefits of such hierarchical content and the advantages of learning over manual construction. After this, I consider alternative paradigms for encoding and acquiring plan expertise before turning to hierarchical task networks. I specify the inputs to HTN learners and three subproblems they must address: identifying hierarchical structure, unifying method heads, and finding method conditions. Finally, I pose seven challenges the community should pursue so that techniques for learning HTNs can reach their full potential.

This paper poses the challenge of developing and evaluating integrated systems for computational scientific discovery. We note some distinguishing characteristics of discovery tasks, examine eight component abilities, review previous successes at partial integration, and consider hurdles the AI research community must leap to transform the vision for integrated discovery into reality. In closing, we discuss promising scientific domains in which to test such computational artifacts.

In this paper, I pose a major challenge for AI researchers: to develop systems that learn in a human-like manner. I briefly review the history of machine learning, noting that early work made close contact with results from cognitive psychology but that this is no longer the case. I identify seven characteristics of human behavior that, if reproduced, would offer better ways to acquire expertise than statistical induction over massive training sets. I illustrate these points with two domains - mathematics and driving - where people are effective learners and review systems that address them. In closing, I suggest ways to encourage more research on human-like learning.

In this paper, I pose a new research challenge – to develop intelligent agents that exhibit radical autonomy by responding to sudden, long-term changes in their environments. I illustrate this idea with examples, identify abilities that support it, and argue that, although each ability has been studied in isolation, they have not been combined into integrated systems. In addition, I propose a framework for characterizing environments in which goal-directed physical agents operate, along with specifying the ways in which those environments can change over time. In closing, I outline some approaches to the empirical study of such open-world learning.

Modern introductory courses on AI do not train students to create intelligent systems or provide broad coverage of this complex field. In this paper, we identify problems with common approaches to teaching artificial intelligence and suggest alternative principles that courses should adopt instead. We illustrate these principles in a proposed course that teaches students not only about component methods, such as pattern matching and decision making, but also about their combination into higher-level abilities for reasoning, sequential control, plan generation, and integrated intelligent agents. We also present a curriculum that instantiates this organization, including sample programming exercises and a project that requires system integration. Participants also gain experience building knowledge-based agents that use their software to produce intelligent behavior.

In this paper, we pose a new challenge for AI researchers – to develop intelligent systems that support justified agency. We illustrate this ability with examples and relate it to two more basic topics that are receiving increased attention – agents that explain their decisions and ones that follow societal norms. In each case, we describe the target abilities, consider design alternatives, note some open questions, and review prior research. After this, we return to justified agency, offering a hypothesis about its relation to explanatory and normative behavior. We conclude by proposing testbeds and experiments to evaluate this empirical claim and encouraging other researchers to contribute to this crucial area.

Research on cognitive architectures attempts to develop unified theories of the mind. This paradigm incorporates many ideas from other parts of AI, but it differs enough in its aims and methods that it merits separate treatment. In this paper, we review the notion of cognitive architectures and some recurring themes in their study. Next we examine the substantial progress made by the subfield over the past 40 years, after which we turn to some topics that have received little attention and that pose challenges for the research community.

Inductive process modeling involves the construction of explanatory accounts for multivariate time series. As typically specified, background knowledge is available in the form of generic processes that serve as the building blocks for candidate model structures. In this paper, we present a more flexible approach that, when available processes are insufficient to construct an acceptable model, automatically produces new generic processes that let it complete the task. We describe FPM, a system that implements this idea by composing knowledge about algebraic rate expressions and about conceptual processes like predation and remineralization in ecology. We demonstrate empirically FPM's ability to construct new generic processes when necessary and to transfer them later to new modeling tasks. We also compare its failure-driven approach with a naive scheme that generates all possible processes at the outset. We conclude by discussing prior work on equation discovery and model construction, along with plans for additional research.

This paper presents a novel approach to inductive process modeling, the task of constructing a quantitative account of dynamical behavior from time-series data and background knowledge. We review earlier work on this topic, noting its reliance on methods that evaluate entire model structures and use repeated simulation to estimate parameters, which together make severe computational demands. In response, we present an alternative method for process model induction that assumes each process has a rate, that this rate is determined by an algebraic expression, and that changes due to a process are directly proportionalto its rate. We describe RPM, an implemented system that incorporates these ideas, and we report analyses and experiments that suggest it scales well to complex domains and data sets. In closing, we discuss related research and outline ways to extend the framework.

In this paper, we discuss a computational approach to the cognitivetask of social planning. First, we specify a class of planningproblems that involve an agent who attempts to achieve its goalsby altering other agents' mental states. Next, we describe SFPS,a flexible problem solver that generates social plans of this sort,including ones that include deception and reasoning about otheragents' beliefs. We report the results for experiments on socialscenarios that involve different levels of sophistication and thatdemonstrate both SFPS's capabilities and the sources of its power.Finally, we discuss how our approach to social planning has beeninformed by earlier work in the area and propose directions foradditional research on the topic.

Scientists use two forms of knowledge in the construction ofexplanatory models: generalized entities and processes that relatethem; and constraints that specify acceptable combinations of thesecomponents. Previous research on inductive process modeling, whichconstructs models from knowledge and time-series data, has relied onhandcrafted constraints. In this paper, we report an approach todiscovering such constraints from a set of models that have beenranked according to their error on observations. Our approach adaptsinductive techniques for supervised learning to identify processcombinations that characterize accurate models. We evaluate themethod's ability to reconstruct known constraints and to generalizewell to other modeling tasks in the same domain. Experiments with synthetic data indicate that the approach can successfully reconstructknown modeling constraints. Another study using natural data suggests that transferring constraints acquired from one modeling scenario to another within the same domain considerably reduces the amount of search for candidate model structures while retaining the most accurate ones.

Quantitative modeling plays a key role in the natural sciences, and systems that address the task of inductive process modeling can assist researchers in explaining their data. In the past, such systems have been limited to data sets that recorded change over time, but many interesting problems involve both spatial and temporal dynamics. To meet this challenge, we introduce SCISM, an integrated intelligent system which solves the task of inducing process models that account for spatial and temporal variation. We also integrate SCISM with a constraint learning method to reduce computation during induction. Applications to ecological modeling demonstrate that each system fares well on the task, but that the enhanced system does so much faster than the baseline version.

We introduce relational temporal difference learning as an effective approach to solving multi-agent Markov decision problems with large state spaces. Our algorithm uses temporal difference reinforcement to learn a distributed value function represented over a conceptual hierarchy of relational predicates. We present experiments using two domains from the General Game Playing repository, in which we observe that our system achieves higher learning rates than non-relational methods. We also discuss related work and directions for future research.

Knowledge-based planning methods offer benefits over classical techniques, but they are time consuming and costly to construct. There has been research on learning plan knowledge from search, but this can take substantial computer time and may even fail to find solutions on complex tasks. Here we describe another approach that observes sequences of operators taken from expert solutions to problems and learns hierarchical task networks from them. The method has similarities to previous algorithms for explanation-based learning, but differs in its ability to acquire hierarchical structures and in the generality of learned conditions. These increase the method's capability to transfer learned knowledge to other problems and supports the acquisition of recursive procedures. After presenting the learning algorithm, we report experiments that compare its abilities to other techniques on two planning domains. In closing, we review related work and directions for future research.

In this paper, we propose a new representation for physical control -- teleoreactive logic programs -- along with an interpreter that uses them to achieve goals. In addition, we present a new learning method that acquires recursive forms of these structures from traces of successful problem solving. We report experiments in three different domains that demonstrate the generality of this approach. In closing, we review related work on learning complex skills and discuss directions for future research on this topic.

In this paper, we review the paradigm of inductive process modeling, which uses background knowledge about possible component processes to construct quantitative models of dynamical systems. We note that previous methods for this task tend to overfit the training data, which suggests ensemble learning as a likely response. However, such techniques combine models in ways that reduce comprehensibility, making their output much less accessible to domain scientists. As an alternative, we introduce a new approach that induces a set of process models from di#erent samples of the training data and uses them to guide a final search through the space of model structures. Experiments with synthetic and natural data suggest this method reduces error and decreases the chance of including unnecessary processes in the model.

In this paper we describe ICARUS, an adaptive architecture for intelligent physical agents. We contrast the framework's assumptions with those of earlier architectures, taking examples from an in-city driving task to illustrate our points. Key differences include: primacy of perception and action over problem solving, separate memories for categories and skills, a hierarchical organization on both memories, strong correspondence between long-term and short-term structures, and cumulative learning of skill hierarchies. We support claims for ICARUS' generality by reporting our experience with driving and three other domains. In closing, we discuss limitations of the current architecture and propose extensions that would remedy them.

This paper describes Eureka, a problem-solving architecture that operates under strong constraints on its memory and processes. Most significantly, Eureka does not assume free access to its entire long-term memory. That is, failures in problem solving may arise not only from missing knowledge, but from the (possibly temporary) inability to retrieve appropriate existing knowledge from memory. Additionally, the architecture does not include systematic backtracking to recover from fruitless search paths. These constraints significantly impact Eureka's design. Humans are also subject to such constraints, but are able to overcome them to solve problems effectively. In Eureka's design, we have attempted to minimize the number of additional architectural commitments, while staying faithful to the memory constraints. Even under such minimal commitments, Eureka provides a qualitative account of the primary types of learning reported in the literature on human problem solving. Further commitments to the architecture would refine the details in the model, but the approach we have taken de-emphasizes highly detailed modeling to get at general root causes of the observed regularities. Making minimal additional commitments to Eureka's design strengthens the case that many regularities in human learning and problem solving are entailments of the need to handle imperfect memory.

Searching for and making decisions about information is becoming increasingly difficult as the amount of information and number of choices increases. Recommendation systems help users find items of interest of a particular type, such as movies or restaurants, but are still somewhat awkward to use. Our solution is to take advantage of the complementary strengths of personalized recommendation systems and dialogue systems, creating personalized aides. We present a system -- the Adaptive Place Advisor -- that treats item selection as an interactive, conversational process, with the program inquiring about item attributes and the user responding. Individual, long-term user preferences are unobtrusively obtained in the course of normal recommendation dialogues and used to direct future conversations with the same user. We present a novel user model that influences both item search and the questions asked during a conversation. We demonstrate the effectiveness of our system in significantly reducing the time and number of interactions required to find a satisfactory item, as compared to a control group of users interacting with a non-adaptive version of the system.

The Stock Tracker is an adaptive recommendation system for trading stocks that automatically acquires content-based models of user preferences to tailor its buy and sell advice. The system incorporates an efficient algorithm that exploits the fixed structure of user models and relies on unobtrusive data-gathering techniques. In this paper, we describe our approach to personalized recommendation and its implementation in this domain. We also discuss experiments that evaluate the system's behavior on both human subjects and synthetic users. The results suggest that the Stock Tracker can rapidly adapt its advice to different types of users

The Stock Tracker is a personalized recommendation system for trading stocks. The system tailors its buy, sell, and hold recommendations to individual users through automatically acquired content-based models of user preferences. It relies on data gathered unobtrusively during the natural course of interacting with a user.

Most AI research on scientific model construction aims to automate this process using discovery techniques. In contrast, we describe an interactive environment for model construction that lets the user construct, edit, and visualize scientific models, use them to make predictions, and call on discovery methods to revise them in ways that better fit the available data. The environment relies on a new formalism that embeds mathematical equations, which are familiar to many scientists, within distinct processes, which can encode background knowledge used to constrain model revision. We report initial studies on ecosystem modeling that suggest this environment is more effective than earlier approaches and more transparent to users. In closing, we discuss related work on modeling environments and model revision, then suggest directions for future research.

Clustering algorithms have become increasingly important in handling and analyzing data. Considerable work has been done in devising effective but increasingly specific clustering algorithms. In contrast, we have developed a generalized framework that accommodates diverse clustering algorithms in a systematic way. This framework views clustering as a general process of iterative optimization that includes modules for supervised learning and instance assignment. The framework has also suggested several novel clustering methods. In this paper, we investigate experimentally the efficacy of these algorithms and test some hypotheses about the relation between such unsupervised techniques and the supervised methods embedded in them.

Many advanced safety and navigation applications in vehicles require accurate, detailed digital maps, but manual lane measurements are expensive and time-consuming, making automated techniques desirable.This paper describes a data-mining approach to map refinement, using position traces that come from Global Positioning System receivers with differential corrections.The computed lane models enable safety applications, such as lanekeeping, and convenience applications, such as lane-changing advice.Experiments show that, starting from a baseline map that is commercially available, our lane models predict a vehicle's lane with high accuracy from a small number of passes over a particular road segment.Multiple position traces are a powerful new source of data that enables cheap, automated methods of inducing lane models, as well as other geographic knowledge, like traffic signals and elevations, and potentially impacts any geographic information system with a need to relate to actual behavior.Keywords: Background knowledge, noisy data, incremental algorithms, implementation and use of KDD systems, case studies, evaluating knowledge and potential discoveries.'The GPS receivers used in this study are generally accurate to between 1 and 2 meters, whereas road lanes are about 3 to 4 meters wide.Pemissjon to make digital or hard copies of all or part of this work fol personal or classroom use is granted without fee provided that cwics are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page.To CWY otherwise, to republish, to post on seners or to redistrihutc 10 Ms. requires prior specific permission and/or a fee.

We present the results of an empirical study in which we evaluated cost-sensitive learning algorithms on a rooftop detection task, which is one level of processing in a building detection system. Specifically, we investigated how well machine learning methods generalized to unseen images that differed in location and in aspect. For the purpose of comparison, we included in our evaluation a handcrafted linear classifier, which is the selection heuristic currently used in the building detection system. ROC analysis showed that, when generalizing to unseen images that differed in location and aspect, a naive Bayesian classifier outperformed nearest neighbor and the handcrafted solution.

In this paper we present a computational theory of human motor performance and learning. The theory is implemented as a running AI system called MAGGIE. Given a description of a desired movement as input, the system generates simulated motor behavior as output. The theory states mat skills are encoded as motor schemas, which specify the positions and velocities of a limb at selected points in time. Moreover, there exist two natural representations for such knowledge; viewer‐centered schemas describe visually perceived behavior, arid joint‐centered schemas are used to generate behavior. When the model acts upon these two representational formats, they exhibit quite different behavioral characteristics. MAGGIE performs the desired movement within a feedback control paradigm, monitoring for errors and correcting them when it detects them. Learning involves improving the joint‐centered schema over many practice trials; this reduces the need for monitoring. The model accounts for a number of well‐documented motor phenomena, including the speed‐accuracy trade‐off and the gradual improvement in performance with practice. It also makes several testable predictions. We close with a discussion of the theory's strengths and weaknesses, along with directions for future research.

Abstract The field of machine learning strives to develop methods and techniques to automate the acquisition of new information, new skills, and new ways of organizing existing information. This article reviews the major approaches to machine learning in symbolic domains, illustrated with occasional paradigmatic examples.

AMBER is a model of first language acquisition that improves its performance through a process of error recovery.The model is implemented as an adaptive production system that introduces new condition-action rules on the basis of experience.AMBER starts with the ability to say only one word at a time, but adds rules for ordering goals and producing grammatical morphemes, based on comparisons between predicted and observed sentences.The morpheme rules may be overly general and lead to errors of commission; such errors evoke a discrimination process, producing more conservative rules with additional conditions.The system's performance improves gradually, since rules must be relearned many times before they are used.AMBER'S learning mechanisms account for some of the major developments observed in children's early speech.